This is a technical report on training a small transformer from scratch in 1.5 hours on a single GPU (RTX 5090) to solve ARC-1 puzzles, achieving 44% accuracy on the public evaluation set. The core method: tokenize input-output grid pairs, train a transformer autoregressively at test time on both training and evaluation puzzle inputs (labels hidden), use 3D positional embeddings and per-task learned embeddings, and apply color/rotation augmentations with majority voting on outputs. Key practical claims: the approach costs ~67 cents lifetime compute, matches performance of prior work (TRM/HRM) without recursive loops, and ablations show positional encoding and per-task embeddings drive most gains. The author argues this contradicts claims that recursion or massive synthetic data are necessary, and critiques the current ARC leaderboard for conflating different evaluation regimes (offline pretraining vs. test-time training, per-task cost accounting). The post heavily defends transductive test-time training as valid metalearning practice and pushes back on viral criticism from prominent researchers who claimed this approach violates testing policy.